Spaces:
Running
Running
File size: 10,763 Bytes
b0b140b 962ad43 b0b140b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | """Observation β prompt rendering + LLM response β action parsing.
Keeps prompt format aligned with Appendix A of LANDSCAPEFORGE_DESIGN.md while
trimming obs fields that bloat tokens (e.g. full trajectories get summarised).
"""
from __future__ import annotations
import json
import re
from typing import Any
try:
from .models import LandscapeforgeAction, LandscapeforgeObservation
except ImportError: # flat layout (HF Space container)
from models import LandscapeforgeAction, LandscapeforgeObservation # type: ignore
SYSTEM = """You are OptCoder. You will design an optimization algorithm for a
hidden landscape f: R^n β R by iteratively: running reference optimizers to
observe their behaviour, writing candidate `Optimizer` classes and seeing how
they perform, inspecting past drafts to diagnose failures, and committing when
you are satisfied.
How the episode ends:
- When you call `commit`, the env runs the full arena evaluation
(10 seeds Γ 200 steps) on your MOST RECENT draft and that becomes your
reward. This is the normal, preferred way to finish.
- If you never call `commit`, when your budget runs out the env will
automatically do the same thing β evaluate your most recent draft.
Your last draft is always what gets evaluated, whether you commit
explicitly or the budget runs out.
- So: make sure your last draft is the one you actually want evaluated.
If you improve a draft then change your mind, re-submit the good one
before ending the episode.
A typical good episode is ~4 turns:
draft β (maybe) inspect β (maybe) refine β commit.
Reply with a single JSON object β nothing else, no prose, no markdown.
JSON formatting rules (important, models frequently get this wrong):
- All strings use standard JSON double-quotes: "like this"
- Do NOT use Python triple-quoted strings \"\"\"...\"\"\" β they are NOT valid JSON
- For multi-line code, escape newlines as \\n inside the string value:
{"kind": "draft", "code": "class Optimizer:\\n def __init__(self, dim): ..."}
""".strip()
ACTION_SPEC = """
Available actions (cost charged against your budget):
run_baseline (cost 2) Run a reference optimizer on the hidden landscape.
JSON: {"kind": "run_baseline", "baseline_name": "sgd"|"momentum"|"adam"|"lbfgs"}
Returns a 30-step trajectory (x_t, f_t, grad_norm_t). Source code not revealed.
draft (cost 2) Submit a full Optimizer class; env auto-tests it.
JSON: {"kind": "draft", "code": "<python source>"}
The code MUST be a standalone class with no base class:
class Optimizer:
def __init__(self, dim):
...
def step(self, x, f_val, grad):
...
return x_new
Rules:
- Top-level line must be exactly: class Optimizer:
(no parent class β BaseOptimizer, nn.Module, object, etc. do NOT exist)
- Use only numpy as `np` and math β both pre-injected; DO NOT write import lines
- step(x, f_val, grad) must return a numpy array of shape (dim,)
- No I/O, no globals, no file operations
- Only the class definition is kept; demo code at module level is stripped
inspect (cost 1) Zoom into a prior draft's per-step behaviour.
JSON: {"kind": "inspect", "draft_idx": 0, "step_range_start": 10, "step_range_end": 20}
Returns per-step (x, f, grad, update_norm, step_size_eff).
commit (cost 0) Evaluate your most recent draft on the full arena.
JSON: {"kind": "commit"}
Preferred way to end the episode. Call it when you have a draft you
trust. If you don't call it, budget exhaustion triggers the same
evaluation on whatever your latest draft is β so your most recent
draft should always be your best one. Committing explicitly just
ends the episode sooner.
""".strip()
def render_observation(obs: LandscapeforgeObservation) -> str:
"""Turn an Observation into a compact prompt-friendly state summary."""
lines: list[str] = []
lines.append(f"Landscape: {obs.landscape_description}")
lines.append(f"Dim: {obs.dim}")
lines.append(f"Structural hints:")
for k, v in (obs.structural_hints or {}).items():
lines.append(f" {k}: {_fmt(v)}")
lines.append(f"Budget remaining: {obs.budget_remaining}")
if obs.baseline_history:
lines.append("\nBaseline runs (diagnostic trajectories):")
for i, b in enumerate(obs.baseline_history):
summary = _summarise_trajectory(b.get("trajectory", []))
lines.append(f" [{i}] {b['name']}: {summary}")
if obs.draft_history:
lines.append("\nDraft history:")
for i, d in enumerate(obs.draft_history):
if d.get("compile_error"):
lines.append(f" [{i}] COMPILE ERROR: {d['compile_error']}")
else:
s = d["summary"] or {}
status = "CONVERGED" if s.get("converged") else (
"DIVERGED" if s.get("diverged") else "partial"
)
lines.append(
f" [{i}] {status} | initial_f={_fmt(s.get('initial_f'))} "
f"final_f={_fmt(s.get('final_f'))} "
f"step_of_min={s.get('step_of_min')}"
)
code = d.get("code") or ""
lines.append(" code:")
for cl in code.splitlines()[:40]: # first 40 lines only
lines.append(f" {cl}")
if obs.inspect_requests:
lines.append("\nInspect results:")
for r in obs.inspect_requests:
detail = r.get("detail") or []
lines.append(
f" draft={r.get('draft_idx')} range={r.get('step_range')} "
f"({len(detail)} steps)"
)
for d in detail[:8]: # first 8 of the slice
lines.append(
f" t={d.get('t'):>3} f={_fmt(d.get('f'))} "
f"|g|={_fmt(d.get('grad_norm'))} "
f"|Ξx|={_fmt(d.get('update_norm'))} "
f"Ξ·_eff={_fmt(d.get('step_size_eff'))}"
)
if obs.current_draft:
lines.append(f"\nCurrent draft ({len(obs.current_draft)} chars) β will be evaluated on commit.")
if obs.last_action_kind:
lines.append(f"\nLast action: {obs.last_action_kind}")
feedback = (obs.last_action_result or {}).get("feedback")
if feedback:
parts = ", ".join(f"{k}={_fmt(v)}" for k, v in feedback.items())
lines.append(f"Step feedback: {parts} "
"(signals for your reasoning; not added to final reward)")
return "\n".join(lines)
def build_prompt(obs: LandscapeforgeObservation) -> list[dict]:
"""Return OpenAI-style messages list for the chat completions endpoint."""
state_text = render_observation(obs)
return [
{"role": "system", "content": SYSTEM},
{"role": "user", "content": f"{ACTION_SPEC}\n\nCurrent state:\n{state_text}\n\n"
"Reply with a single JSON object for your next action."},
]
# ---------- response β action ----------
_JSON_RE = re.compile(r"\{.*\}", re.DOTALL)
def parse_action(response_text: str) -> LandscapeforgeAction:
"""Extract the first JSON object from the LLM's reply and build an Action.
Accepts code-fenced JSON, raw JSON, and JSON embedded in prose. Tolerates
the common LLM failure mode of emitting unescaped newlines / tabs inside
string values (especially for the `code` field of a `draft` action).
Raises ValueError if no parseable object is found.
"""
text = response_text.strip()
if text.startswith("```"):
text = re.sub(r"^```(?:json)?\n?", "", text)
text = re.sub(r"\n?```\s*$", "", text)
match = _JSON_RE.search(text)
if not match:
raise ValueError(f"No JSON object in response: {response_text[:200]!r}")
raw_json = match.group(0)
# First pass: strict.
try:
data = json.loads(raw_json)
except json.JSONDecodeError:
# Second pass: escape raw control chars inside string literals.
fixed = _escape_string_controls(raw_json)
try:
data = json.loads(fixed)
except json.JSONDecodeError as e:
raise ValueError(f"Invalid JSON even after control-char fix: {e}; "
f"raw: {raw_json[:200]!r}") from e
if "kind" not in data:
raise ValueError(f"Missing `kind`: {data}")
return LandscapeforgeAction(**data)
def _escape_string_controls(s: str) -> str:
"""Escape raw newlines, carriage returns, and tabs inside JSON string literals.
Walks character-by-character tracking whether we're inside a double-quoted
string, and replaces raw control chars with their escaped forms. Handles
the common case: `"code": "class Optimizer:\\n def __init__..."` where the
LLM emitted literal newlines.
"""
out: list[str] = []
in_string = False
escape_next = False
for ch in s:
if escape_next:
out.append(ch)
escape_next = False
continue
if ch == "\\":
out.append(ch)
escape_next = True
continue
if ch == '"':
in_string = not in_string
out.append(ch)
continue
if in_string:
if ch == "\n":
out.append("\\n"); continue
if ch == "\r":
out.append("\\r"); continue
if ch == "\t":
out.append("\\t"); continue
out.append(ch)
return "".join(out)
# ---------- helpers ----------
def _fmt(v: Any) -> str:
if v is None:
return "None"
if isinstance(v, float):
if abs(v) < 1e-4 or abs(v) >= 1e4:
return f"{v:.3e}"
return f"{v:.4f}"
if isinstance(v, list):
if len(v) <= 4:
return "[" + ", ".join(_fmt(x) for x in v) + "]"
return f"[{_fmt(v[0])}, {_fmt(v[1])}, ..., {_fmt(v[-1])}] (len={len(v)})"
return str(v)
def _summarise_trajectory(traj: list[dict]) -> str:
"""Condense a 30-step baseline trajectory to head/tail snapshots."""
finite = [s for s in traj if s.get("f") is not None]
if not finite:
return "diverged immediately"
head = finite[0]
mid = finite[len(finite) // 2] if len(finite) > 2 else finite[-1]
tail = finite[-1]
diverged_mark = " (DIVERGED)" if len(finite) < len(traj) else ""
return (f"t=0: f={_fmt(head['f'])}, |g|={_fmt(head['grad_norm'])} "
f"β t={mid['t']}: f={_fmt(mid['f'])} "
f"β t={tail['t']}: f={_fmt(tail['f'])}{diverged_mark}")
|